CUDA driver too old — update path + minimum versions for 2026
If PyTorch / vLLM / CUDA app errors with 'driver version insufficient,' your NVIDIA driver predates the CUDA runtime. Driver 555+ supports CUDA 12.4 (the 2026 standard). Update via nvidia.com or distro.
Diagnostic order — most likely first
Driver < 550 with CUDA 12.4 toolkit / wheel
`nvidia-smi` upper-right shows e.g. 'CUDA Version: 12.0'. PyTorch 2.5+ wants 12.4. Error: 'CUDA driver version is insufficient for CUDA runtime version.'
Update driver. Linux: `sudo apt install nvidia-driver-555` (Ubuntu) or distro equivalent. Windows: download Game Ready Driver 555+ from nvidia.com. Reboot. Verify `nvidia-smi` shows CUDA 12.4+.
Driver < 535 (predates WSL CUDA support)
WSL2 user. `nvidia-smi` inside WSL fails. Host driver is too old for WSL passthrough.
Update Windows host driver to 535+ (preferably 555+). WSL passthrough requires inbox CUDA support which only ships in 535+. After update, reboot Windows.
Pinned PyTorch version requires newer CUDA than your driver
PyTorch 2.5+ with cu124 wheel, but driver maxes out at CUDA 12.1. Skew is on the wheel side.
Either update driver (preferred) or pin PyTorch to a wheel matching your driver: `pip install torch --index-url https://download.pytorch.org/whl/cu118` (for older drivers) or cu121.
Linux kernel module didn't reload after driver upgrade
Driver was updated but `nvidia-smi` still shows old version, or fails entirely. `lsmod | grep nvidia` shows old module.
Reboot. Or unload + reload: `sudo modprobe -r nvidia && sudo modprobe nvidia`. On Ubuntu after dkms rebuild: reboot is the safest path.
Datacenter GPU driver branch (R535/R570) skew
On a server with a datacenter GPU (A100, H100), the LTSB driver is on a different branch than consumer. CUDA 12.4 needs R555-server or R560-server.
Use NVIDIA's datacenter driver matrix to find the right version. Don't mix consumer and datacenter drivers on the same machine.
Frequently asked questions
Does updating NVIDIA driver break anything?
Almost never on Linux/Windows desktop. Drivers are highly backwards-compatible — newer drivers run older CUDA apps. Rare failures: bleeding-edge driver + unmaintained third-party tool. For production servers, test before mass-rollout.
Should I use the latest driver or a stable one?
For local AI: latest stable (555+ in 2026). Avoid 'beta' or 'developer preview' branches unless you specifically need a fix from them. NVIDIA's quarterly stable releases are well-tested.
Can I have CUDA 12.4 toolkit with an older driver?
No — CUDA 12.4 toolkit at runtime requires driver 555+ to load. The toolkit can be installed but apps using it will error at runtime. The driver is the floor; the toolkit is the ceiling.
Related troubleshooting
When PyTorch / vLLM / a CUDA app errors on 'CUDA driver version is insufficient' or 'no kernel image,' the host driver is too old (or sometimes too new) for the installed toolkit. Read nvidia-smi's max-CUDA, match it.
PyTorch falsely reporting no CUDA is the most common Python ML setup failure. The cause is almost always: wrong PyTorch wheel for your CUDA version, or a CPU-only build accidentally installed.
WSL2 doesn't pass the GPU through unless the host driver is right and the kernel is current. Here's the install order that actually works in 2026, and how to confirm passthrough is live before you waste an afternoon.
When the fix is hardware
A surprising fraction of troubleshooting tickets resolve to: this card doesn't have enough VRAM for what you're asking it to do. If you're hitting OOM after every reasonable fix, or your GPU genuinely can't fit the model you need, it's upgrade time: